Preface
I installed pytorch in the Alibaba Cloud docker virtual environment and went through a lot of pitfalls to record it.
This time I want to build an environment under Alibaba's domestic arm environment python3.8 + pytorch1.10.2. Since torch has been updated to 2.0, the old version of the arm environment has to be installed again.
Environmental Statement
Server environment
Ali Cloud
・实例: 1核 2G ecs.c8y ecs-7
・I/O 优化实例: I/O 优化实例
・系统盘: 增强型SSD云盘 /dev/xvda 40GB 模块属性 PL0
・操作系统: Alibaba Cloud Linux 3.2104 LTS 64位 ARM版 Linux 64位
target environment
python==3.8.16
torch==1.10.2
detailed version
The system and CPU environment are as follows
cat /proc/version
Linux version 5.10.134-13.1.al8.aarch64 ([email protected]) (gcc (GCC) 10.2.1 20200825 (Alibaba 10.2.1-3 2.32), GNU ld version 2.35-12.2.al8) #1 SMP Mon Feb 6 07:07:12 UTC 2023
uname -a
Linux 810a9ecee45c 5.10.134-13.1.al8.aarch64 #1 SMP Mon Feb 6 07:07:12 UTC 2023 aarch64 GNU/Linux
For commands to check the Linux kernel version, please refer to How to Check the Linux System Version?
1. Pull the docker image
Here select the python3.8 environment image officially provided by docker, address
Image pull command
docker pull python:3.8.16-buster
# 这个镜像需要手动添加yum。使用下面语句更新apt-get，安装yum
apt-get -y update
apt-get install yum
For how to use docker, please refer to the Docker Concise Usage Guide.
2. Select the corresponding version offline package.whl
Online installation
Because it is no longer possible to select the corresponding version through online installation, offline installation is used. For newer versions, you can check the online installation version first:
pip install torch==
Generates error: No version available
pip install torch==
Looking in indexes: http://mirrors.aliyun.com/pypi/simple/
ERROR: Could not find a version that satisfies the requirement torch== (from versions: none)
ERROR: No matching distribution found for torch==
Pytorch officially does not provide an installation package for the arm platform,
so the installation command provided on the official page cannot be installed successfully.
As long as the CPU is an arm chip, it will not work regardless of windows, linux, maxos.
There are two solutions
: 1. Download the source code and then proceed. Manual compilation (too complicated and time-consuming)
2. Use the version compiled by others in the community. This is relatively simple.
Download offline installation package
Here we directly use doctrine and download the corresponding offline version installation package according to the system version. whl
torch download address in Tsinghua source
. Here we find the installation package from Tsinghua source torch-1.10.2-cp38-cp38-manylinux2014_aarch64.whl
, which cp38represents python 3.8, aarch64which is arm64 .
3. Install offline packages
Here try to install the offline package directly
pip install XXX.whl
#如当前准备安装的 torch1.10.2 的linux安装包
pip install torch-1.10.2-cp38-cp38-manylinux2014_aarch64.whl
produces error: is not a supported wheel on this platform
The whl installation package is not supported by this platform (actually not)
pip install torch-1.10.2-cp38-cp38-manylinux2014_aarch64.whl
Looking in indexes: http://mirrors.aliyun.com/pypi/simple/
ERROR: torch-1.10.2-cp38-cp38-manylinux2014_aarch64.whl is not a supported wheel on this platform.
But in fact, we found that Linux here only has this version of the pytorch installation package, so we have no choice.
Solution: Modify the installation package name
In fact, it’s just that the suffix of the installation package is wrong. The system determines that it recognizes the suffix of the whl file (so stupid).
First query
pip debug --verbose
(The statement is applicable to new versions of pip such as pip 20.0. You need to refer to this for previous pip versions. )
After entering this command, a large amount of information will be displayed:
$ pip debug --verbose
WARNING: This command is only meant for debugging. Do not use this with automation for parsing and getting these details, since the output and options of this command may change without notice.
pip version: pip 23.0.1 from /usr/local/lib/python3.8/site-packages/pip (python 3.8)
sys.version: 3.8.16 (default, Mar 29 2023, 23:56:55)
[GCC 12.2.1 20220924]
sys.executable: /usr/local/bin/python
sys.getdefaultencoding: utf-8
sys.getfilesystemencoding: utf-8
locale.getpreferredencoding: UTF-8
sys.platform: linux
sys.implementation:
name: cpython
'cert' config value: global
REQUESTS_CA_BUNDLE: None
CURL_CA_BUNDLE: None
pip._vendor.certifi.where(): /usr/local/lib/python3.8/site-packages/pip/_vendor/certifi/cacert.pem
pip._vendor.DEBUNDLED: False
vendored library versions:
CacheControl==0.12.11
colorama==0.4.6
distlib==0.3.6
distro==1.8.0
msgpack==1.0.4
packaging==21.3
platformdirs==2.6.2
pyparsing==3.0.9
pyproject-hooks==1.0.0
requests==2.28.2
certifi==2022.12.07
chardet==5.1.0
idna==3.4
urllib3==1.26.14
rich==12.6.0 (Unable to locate actual module version, using vendor.txt specified version)
pygments==2.13.0
typing_extensions==4.4.0 (Unable to locate actual module version, using vendor.txt specified version)
resolvelib==0.8.1
setuptools==44.0.0 (Unable to locate actual module version, using vendor.txt specified version)
six==1.16.0
tenacity==8.1.0 (Unable to locate actual module version, using vendor.txt specified version)
tomli==2.0.1
webencodings==0.5.1 (Unable to locate actual module version, using vendor.txt specified version)
Compatible tags: 87
cp38-cp38-musllinux_1_2_aarch64
cp38-cp38-musllinux_1_1_aarch64
cp38-cp38-musllinux_1_0_aarch64
cp38-cp38-linux_aarch64
cp38-abi3-musllinux_1_2_aarch64
cp38-abi3-musllinux_1_1_aarch64
cp38-abi3-musllinux_1_0_aarch64
cp38-abi3-linux_aarch64
cp38-none-musllinux_1_2_aarch64
cp38-none-musllinux_1_1_aarch64
cp38-none-musllinux_1_0_aarch64
cp38-none-linux_aarch64
cp37-abi3-musllinux_1_2_aarch64
cp37-abi3-musllinux_1_1_aarch64
cp37-abi3-musllinux_1_0_aarch64
cp37-abi3-linux_aarch64
cp36-abi3-musllinux_1_2_aarch64
cp36-abi3-musllinux_1_1_aarch64
cp36-abi3-musllinux_1_0_aarch64
cp36-abi3-linux_aarch64
cp35-abi3-musllinux_1_2_aarch64
cp35-abi3-musllinux_1_1_aarch64
cp35-abi3-musllinux_1_0_aarch64
cp35-abi3-linux_aarch64
cp34-abi3-musllinux_1_2_aarch64
cp34-abi3-musllinux_1_1_aarch64
cp34-abi3-musllinux_1_0_aarch64
cp34-abi3-linux_aarch64
cp33-abi3-musllinux_1_2_aarch64
cp33-abi3-musllinux_1_1_aarch64
cp33-abi3-musllinux_1_0_aarch64
cp33-abi3-linux_aarch64
cp32-abi3-musllinux_1_2_aarch64
cp32-abi3-musllinux_1_1_aarch64
cp32-abi3-musllinux_1_0_aarch64
cp32-abi3-linux_aarch64
py38-none-musllinux_1_2_aarch64
py38-none-musllinux_1_1_aarch64
py38-none-musllinux_1_0_aarch64
py38-none-linux_aarch64
py3-none-musllinux_1_2_aarch64
py3-none-musllinux_1_1_aarch64
py3-none-musllinux_1_0_aarch64
py3-none-linux_aarch64
py37-none-musllinux_1_2_aarch64
py37-none-musllinux_1_1_aarch64
py37-none-musllinux_1_0_aarch64
py37-none-linux_aarch64
py36-none-musllinux_1_2_aarch64
py36-none-musllinux_1_1_aarch64
py36-none-musllinux_1_0_aarch64
py36-none-linux_aarch64
py35-none-musllinux_1_2_aarch64
py35-none-musllinux_1_1_aarch64
py35-none-musllinux_1_0_aarch64
py35-none-linux_aarch64
py34-none-musllinux_1_2_aarch64
py34-none-musllinux_1_1_aarch64
py34-none-musllinux_1_0_aarch64
py34-none-linux_aarch64
py33-none-musllinux_1_2_aarch64
py33-none-musllinux_1_1_aarch64
py33-none-musllinux_1_0_aarch64
py33-none-linux_aarch64
py32-none-musllinux_1_2_aarch64
py32-none-musllinux_1_1_aarch64
py32-none-musllinux_1_0_aarch64
py32-none-linux_aarch64
py31-none-musllinux_1_2_aarch64
py31-none-musllinux_1_1_aarch64
py31-none-musllinux_1_0_aarch64
py31-none-linux_aarch64
py30-none-musllinux_1_2_aarch64
py30-none-musllinux_1_1_aarch64
py30-none-musllinux_1_0_aarch64
py30-none-linux_aarch64
cp38-none-any
py38-none-any
py3-none-any
py37-none-any
py36-none-any
py35-none-any
py34-none-any
py33-none-any
py32-none-any
py31-none-any
py30-none-any
Here we see the words Compatible tags. These are the tags that the current Python version can adapt to, that is, they can match the following file names.
It is found that there is no suffix for our existing installation package, try to modify the suffix of the installation package.
It seems that this suffix is more suitable for cp38-abi3-linux_aarch64.
cp torch-1.10.2-cp38-cp38-manylinux2014_aarch64.whl torch-1.10.2-cp38-abi3-linux_aarch64.whl
Try installing again
pip install torch-1.10.2-cp38-abi3-linux_aarch64.whl
Display the following information
pip install torch-1.10.2-cp38-abi3-linux_aarch64.whl
Looking in indexes: http://mirrors.aliyun.com/pypi/simple/
Processing ./torch-1.10.2-cp38-abi3-linux_aarch64.whl
Collecting typing-extensions
Downloading http://mirrors.aliyun.com/pypi/packages/31/25/5abcd82372d3d4a3932e1fa8c3dbf9efac10cc7c0d16e78467460571b404/typing_extensions-4.5.0-py3-none-any.whl (27 kB)
Installing collected packages: typing-extensions, torch
Successfully installed torch-1.10.2 typing-extensions-4.5.0
Successful installation! scatter flowers
Conclusion
Although it is a pytorch installation method in the arm environment, it can actually be applied to the installation of most python libraries.
References
[1] How to check the Linux system version?
[2] Problem promotion post: Installation of pytorch, torchvision, torchaudio under arm64 and the version correspondence between them
[3] Problem ending post: [2021 new tutorial] Solving is not a supported wheel on this platform - Solving pip has no attribute pep425tags-Solve the problem that old online tutorials are not applicable
Other references:
[4] Torch, torchvision, Python version correspondence and installation of GPU or CPU version of pytorch
[5] Installing pytorch under arm
[6] Installing python's pytorch library under Arm/Linux